From: AAAI Technical Report SS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. Representing ProblemSolving for Tutoring and Critiquing Medical Domains Kathy A. Johnson, Todd R. Johnson, and Jack W. Smith, Jr. Laboratory for Knowledge-Based Medical Systems Division of Medical Informatics The Ohio State University 571 Health Sciences Library 376 W. 10th Ave. Columbus, Ohio 43210 Email: kj@med.ohio-state.edu,tj@med.ohio-state.edu,smith.30@magnus.ohio-state.edu 1 Introduction Medicalproblemsolving often involves large amountsof d~t~, This posesproblemsfor expertssolvingthe problemas well as for studentslearningto do so. Thereare often many different waysto solve a single problem. Wehave been studyingthe constructionof tutoring systemsfor problems involvingthe interpretation of data. Asystemwasbuilt to detect errors in solving the problemof red blood cell antibody identification. The system has at its core a monitoringshell and a representation of problemsolving. Theshell providesa generalmethodfor monitoringthat can be used in any domain for which an appropriate representation can be built. Theshell and representation could also be used as part of a critiquing systemto aid medicalproblemsolving. 2 Representing Tutoring Knowledge Manyrepresentationsfor tutoring havebeenexploredas well as techniquesfor tutoring variousdomains.Giventhat a student has had somebasic training in a problemsolving area, tutoring can be viewedas a process during whicha systemmust1) monitora student for mistakes,2) determine the causeof the mistake,and3) give the studentappropriate informationto correct his mistake.Monitoringfor mistakes maybe as simple as checking the final answer of an arithmetic problem or as complicated as detecting an incorrect sequenceof steps during medicaldiagnosis. An exampleof a medicaltutoring systemtaking this viewof tutoring is Clancey’s GUIDON system which uses the knowledgebase from MYCIN to tutor students in the methodfor doingmedicaldiagnosis (Clancey,1987). Other medicalsystemshave approachedtutoring in a different mannerbest characterized by the term critiquing. Rather than correcting a student’s problem-solvingmethod,these systemspresentrelevantinformationandalternatives to the student. Miller’s VQ-Attending is an exampleof this type of system(Miller, 1986). Foreither tutoring strategy, it is importantto understand the problemsolving taking place. This is obviousfor a systemthat is trying to teach problem-solvingstrategy. * Thisresearchis supportedby NationalHeartLungandBlood Institute grant HL-38776, andNationalLibraryof Medicine grant LM-04298. 74 However,a critiquing system must also knowsomething about the problemsolving in order to present relevant information. In orderto filter all the information availableto aid problemsolving, the systemneeds knowwhatgoals the informationwill be serving. Wehave been studying the domainof red blood cell antibody identification whichis a multi-step problemsolving process similar to medicaldiagnosis. Data from manysources mustbe incorporated to determinethe best answeras to the patient’s antibodies.Thereis a great deal of variability in correct problemsolvingbehavior.Thestudent should be allowedto pursue any correct problem-solving path. In order to build such a system, we needed a representation of problem solving that is capable of encodingall possible correct problem-solving paths. Oneof the mostdifficult aspects of describingflexible problem solving is finding a wayto easily represent alternative methodsto solve a problem. The standard representation for problem solving is a procedural descriptionin whicheachoperator is listed in order along with conditionals and loops. Such descriptions are notoriously inflexible and brittle (Newell, 1990). The ordering of the operatorsmustbe completelyspecified and the justification for choosingoneoperatorbeforeanotheris lost in the encoding. Analternative representation for problem-solvingmethodsis embodiedin the problemspace computation model (PSCM)of NeweU (Newell, 1990). the PSCM, all problemsolvingis viewedas searchfor a goal state in a problemspace. Theproblemspace contains a set of operatorswhichwhenapplied in the appropriatesequence will achieve the goal. Knowledge about whenparticular operators are applicable to a state can be specified independentof knowledgeabout whichoperator to prefer. For example,conditions maybe appropriate for moving BlockXor BlockY(e.g. both are clear). Other kinds knowledge mayindicate that it makesno differencewhichis moved first, or perhapsthat it is better to moveBlockXfirst. Operatorsmaybe encodeddirectly or they maythemselves be set up as a goal in a sub-problem spacewith their ownset of operators andcontrol knowledge.For example,movinga block maybe a primitive operator tied to a motorsystem whileclearing a block mightrequire a subgoalto determine the sequenceof operatorsneededto removethe block on top will be achieved, and the behavior by whichthe function is achieved. The behavior is a sequence of state transitions connectedby subfunctions. In our case, a function is a goal to be achieved,the result is the state of problemsolving after achievingthe goal, there are still preconditions, and wehave a method(or methods) for achieving the goal instead of specific device behavior. The difference between a device behavior and a problem-solvingmethodis ~n the variability of the steps. A device usually behaves in a deterministic mannersince a device is a physical system that works in a linear fashion (seldom do devices randomly execute functions unless they are designed to do so). The methodfor problemsolving should allow for variability in sequenceand is represented as a set of subgoals with whatever control knowledgeis necessary (or desired). Goal: Identify-Allo-Antibodies It: not auto-antibodies, desiredallo-antibodiesidentified To-Make:Complete,Consistent, Parsimonious explanationfor data By: Methodl Method:Methodl Goals: Make-abstract-hypotheses, Match-hypothesis-toantigram,Rule-out, Explaindatum Control:Rule-outis better than Explaindatum Otheroperatorsare indifferent Goal: Make-abslract-hypotheses If: datumneedsto be explained To-Make:antibody a explains datum By: Group-by-rows, Group-by-columns, Group-by-likereactions Control:Group-by-rows,Group-by-columns, and Groupby-like-reactions are indifferent. 3 Monitoring Students Using a Functional Representation of Problem Solving Goal: Match-hypothesis-to-antigram If: antibody<a>doesnot havea specificity To-Make: antibody<a>has a specificity By: Method3 The functional representation of problemsolving for red blood cell antibody identification was extracted from a task analysis that was used to develop an expert system(Johnson, et. al. 1991). Thetask analysis wasfacilitated by the use of problemspaces describing generic tasks (Johnson, 1991). portion of the functional representation is shownin Figure 1. A monitoring system (Johnson, 1993) using this functional representation has been built in Soar. It is capable of mappinga simulated student’s intermediate results into the goal that he waspursuingand determiningif there is an error in the student’s reasoning1. Figure 2 showsthe interactions of the componentsof the monitoring shell. Goal: Rule-Out If: desiredlimit-possible-explanations To-Make: antibody<b>ruled-out (i.e. not consideredas part of the explanation) By: Method4 Goal: Explain datum If: datumneedsto be explained To-Make:antibody a explains datum By: Method5 3.1 Monitoring Goals The following are the goals that comprise the monitoring task: ¯ Get-student-result: Get the student’s intermediate conclusion. ¯ Generate-explanations-for-result: Determinethe goals that the student could have been pursuing that would generate such a result ¯ Determine-possible-goals: Determinethe goals that are possible from the previous goal that generate visible results. ¯ Evaluate-validity-of-goal: For a particular goal generated by generate-explanations-for-result, determine whether the goal was valid by checking the preconditions and search-control. Figure 1: Partial modelof goals and methodsfor red blood cell antibodyidentification of it. A description of problemsolving in these terms can provide a more dynamicor flexible description than does an algorithm because operator proposal knowledgeand searchcontrol knowledgecan be sensitive to the current state of problem solving. In addition, alternative methods for achieving goals can be represented as different sets of subgoals and knowledge to choose between alternative methodsmay or maynot be provided. The PSCM allows us to describe flexible problem-solving methods that we need for monitoring a student’s problem solving. To makeuse of this knowledge, however, we need to explicitly represent what each problemspace and operator does. Wehave chosen to represent that knowledge using terminology borrowed from the functional representation developed at Ohio State (Sembugamoorthy and Chandrasekaran, 1986). Previously, the functional representation had been used to describe devices, plans, and computer programs (Allemang, 1990; Chandrasekaran, et al., 1986; Keuneke,1989; Sticklen, 1987). The traditional functional representation consists of a function, whatresult is expected, what conditions mustbe true before the function 1. Thecurrentmonitoris in a versionof Soarwrittenin Lispandis too slowto be usedwithactual students.Additionally,a user interface needsto be developed to capturethe student’sresults. Thedata usedto test the monitorwasrepresentativeof student’sactual results. 75 state and the modelof the student’s state is available for further processing. The general methodfor composingthese goals is: 1. The system scans the problem-solvingsituation that is given to the student to get a preliminary modelof the situation. Tutor Monitor ;Student’sConclusion error indicadon student’sstate Causeof Shell 2. Solvina Model" 3. Student’s Goal: O1 Result: RI . Correction . Correction Student’s Imermediam Conclusions ExternalWorldi.1. }1I I I I I I I II II II II II II II IIIII I II I Intermediate Conclusions Resultsof pro.blem.splving and outside information Figure 2: The high level problem-solving breakdownof the tutor and the monitor ¯ . Determine-explanation: Of the valid explanations (goals) generated, determine which fits the student best. ¯ Evaluate-validity-of-method: Determine whether the conclusion was correct for the goal being pursued. This is generally accomplishedby running the appropriate portion of the expert system and comparingthe result to the student’s intermediate conclusion. These are general goals that rely on having an explicit functional model of problem solving and some way of checking the correctness of a conclusion. That is, they are useful in any system that has those properties. They are incorporated into the expert systems and are used whenthe system has a goal of monitor-student-for-error rather than the expert’s usual domain goal such as identify-alloantibodies in the case of red blood cell antibody identification. Get-student-resuh has domain-specific parts that dependon the particular information available such as menuselections on a screen, mouseclicks in active areas, text input, etc. Likewise, evaluate-validity-of-methodneeds to have access to a knowledge-basedsystem in the domain. Whenthe monitoringshell detects an error, it is noted in its Get-student-resultis set as a goal and it returns the student’s intermediate conclusion. Generate-explanations-for-resuh is used to determine the goals that the student could have been pursuing. Determine-applicable-goals generates a set of goals that are appropriate in the current situation by finding goals whosepreconditions match. Search control is not checkedfor these goals, it is deferred to evaluate-valid. ity-of-goal. Evaluate-validity-of-goal is used for each goal in the set of goals returned from generate-explanations-forresult. If the goal beingtested does not appearin the set of applicable goals generated by determine-applicablegoals, then the goal is markedinvalid and a reason of not-applicable is listed. If there is a match,then search control is checked. If the search control indicates that the goal is not the best one at this time, then the goal is markedinvalid and a reason of wrong-timeis listed. If the search control checks out, then the goal is marked as valid. Determine-explanationis used after all the goals have been checkedfor validity. If no goal is valid, then an error is determinedand the set of goals and reasons for them being invalid is madeavailable for further error diagnosis. If one goal is valid, then it is assumedthat the student is pursuingthis goal even if there wereother invalid goals. If morethan one goal is valid, then one is picked at randomand the others are kept in the student modelas alternative goals. Evaluate-validity-of-method is then used to check the actual value of the student’s conclusion to makesure that it matchesthe correct value. At this point, the expert systemis run to lind the correct value. If it does not match, then an error is determined. If it matches then the student modelis updated with the conclusion. If no errors were indicated, then the process begins again from step 2. . 3.2 Classes of error Thereare several classes of error that the monitoringshell detects. 1. WrongtimingmThe student may be pursuing a goal that is applicable, but is not the best one at this time. That is, the preconditions for the goal were met, but 76 detecting errors in problemsolving. This shell maybe used at the core of a tutoring systemor a critiquing system.It providesthe capability to followa problem-solving path in a flexible manner.Theproblem-solving representationthat is a keypart of the shell providesa goal-orientedindexof methods and could be expanded to also index data relevancy. Future workwill explore both the tutoring and critiquingaspectsof the monitoring shell. search-controlindicated that there wassomethingelse that shouldbe donefirst. 2. Goal not applicable---The student was attempting a goal whosepreconditionswerenot met. 3. Goalnot part of current method--Thestudent wasattemptinga goal not in the current method.Whether the newgoal’spreconditionsweremetor not, this is an error becausethe currentgoal hasn’t beenachievedyet. 4. Bad implementation of goal--The student was attemptinga correct goal, but the result of executingthe goal wasnot correct. It is necessaryto detect these types of errors in any domain. Each error type will have domain-specific explanations. There are also process-specific errors and explanations for reasoning strategies such as abductive assembly.Theseerrors andexplanationswouldbe useful in any domainemployingthe methodof abductiveassembly. Acknowledgments Wethank the AIMgroup for their assistance and comments on this paper and the workit reports. Wealso thank the membersof the Soar communityfor their intellectual andtechnicalsupport.Thisresearchis supported by National Heart Lungand BloodInstitute grant HL38776,and National Library of Medicinegrant LM4)4298. References Allemang, D. T. (1990) UnderstandingProgramsas Devices. Ph.D., TheOhioState University. Chandrasekaran,B. (1986). Generic Tasks in Knowledge-BasedReasoning:High-LevelBuildingBlocksfor Expert SystemDesign.IEEEExpert, 1(3), 23-30. Chandrasekaran,B., Josephson, J., and Keuneke,A. (1986). FunctionalRepresentationsas a Basis for Generating Explanations.In Proceedingsof the 1986IEEEInternational Conferenceon Systems, Manand Cybernetics (pp. 726-731).Atlanta, Georgia:IEEEComputerSociety Press. Clancey,W.J. (1987). Knowledge-Based Tutoring; The GuidonProgram.MITPress. Johnson,K. (1993) Exploiting a FunctionalRepresentation of ProblemSolvingfor ErrorDetectionin Tutoring. Ph.D., OSU. Johnson,K., Johnson,T., Smith, J. W., DeJongh,M., Fischer, O., Amra,N., and Bayazitoglu, A. (1991). RedSoar--A Systemfor RedBloodCell AntibodyIdentification. In Symposiumon ComputerApplications in Medical Care. Washington,D.C. Johnson,T. R. (1991) GenericTasks in the ProblemSpace Paradigm:Building Flexible KnowledgeSystems While UsingTask-LevelConstraints. Ph.D., OSU. Keuneke,A. (1989) MachineUnderstandingof Devices: CausalExplanationsof DiagnosticConclusions.Ph.D., TheOhioState University. Laird, J. E., Newell,A. andRosenbloom, P.S. (1987). SOAR: AnArchitecturefor GeneralIntelligence. Artificial Intelligence, Vol.33(No.1). Newell,A., Yost., G., Laird, J.E., Rosenbloom, P.S. and Altmann, E. (1990 September24-26). Formulating The ProblemSpaceComputationalModel.In Proceedingsof the 25th AnniversarySymposium,School of ComputerScience, CarnegieMellonUniversity. Pittsburgh, PA. Miller, P. L. (1986)ExpertCritiquingSystems:Practice-Based MedicalConsultation by Computer.NewYork: Springer-Verlag. Sembugamoorthy, V., and Chandrasekaran,B. (1986). Functional Representationof Devicesand Compilationof Diagnostic ProblemSolving Systems.In J. Kolodner&C. Reisbeck(Eds.), Experience, Memory and Reasoning(pp. 47-73). LawrenceErlbaumand Associates. S ticklen, J. (1987)MDX2: An IntegratedMedicalDiagnostic System. Ph.D. OSU. 4 CritiquingUsinga Functional Representationof ProblemSolving Wehave shownhowa functional representation of problemsolving can be used to monitora student for errors (an essential subtaskof tutoring). Thesamemonitormay used in the context of a critiquing system. A critiquing systemprovidesa personwith supportfor problemsolving. It may indicate possible omissions of steps and inconsistencies in conclusions. It can provide access to descriptionsof alternative methodsto achievea particular step. It can also help to organizethe informationpresented to a person.Thesefunctionsare possibleonlyif the system is able to determinethe goal of the person solving the problem. Themonitoringshell andfunctional representationcould be used in a critiquing systemas follows. Themonitoring shell infers the problem solver’s goal. If the problem solver has omitteda step, the monitordetects this andputs up a warning.Asthe personsolves the problem,the monitoralso compares the person’sanswerto its expert systemanswer.If these are foundto be significantly different, the system warnsthe person. Since the critiquing systemcontains a description of the problemsolving organizedaroundgoals andmethods,it has easy access to a list of methods.These couldbe madeavailableuponrequest. Finally, the problemsolving methodsin the system makereference to certain types of data. Enhancement of the functional representation wouldprovide a goal-oriented index to the relevant data. Thiscouldthen be usedto filter the data for displayto the humanproblemsolver. 5 Conclusion Medical problem solving is a complexprocess often involving large amountsof data. Solving such problems createsdifficulties for expertssolvingthe problem as wellas for students learning to do so. Wehavebuilt a shell for 77